{ "artifacts": [ { "file": "SmolLM3-3B.litertlm", "sha256": "f9aa844c27f5ccfae9cb701c7ca63f2faba501ed187aacc85044b6200df15a37", "size_mb": 2965.286 } ], "benchmarks": [], "conversion": { "command": "TODO (not stated in the README)", "quantization": "TODO — the README describes this ~3.1 GB file only as 'the repo also carries SmolLM3-3B.litertlm, gated on the Galaxy S26 GPU'; the id suffix -int8 is the device-run lane's, not a README statement", "tool": "litert-torch export_hf (README Conversion, written for the repo's q4 file; SmolLM3ForCausalLM on the existing converter, NoPE attention lowers to generic ops)", "tool_version": "TODO (the README names the converter but not its version)" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 14.41, "delegated_ops": 1548, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-cl-pinned", "os_build": "Android 16", "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "VERBOSE: Replacing 1382 out of 1382 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (decode).", "VERBOSE: Replacing 1548 out of 1548 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_256).", "results block: prefill=312.27 decode=14.41 tokens/s" ], "failure_class": null, "full_delegation": true, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 295.0, "init_s": 4.91261, "prefill_tokens": 211.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 312.27, "provenance": "measured", "runs": true, "total_ops": 1548, "ttft_ms": 750.0 }, "source": "data/device_runs/0.16.0/2026-08-24/smollm3-3b-int8__galaxy-s26.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-01", "decode_tokens_per_s": 1.97, "delegated_ops": null, "env": { "device": "Raspberry Pi 5 Model B Rev 1.1", "machine_label": "raspberry-pi-5-cooled-52c", "os_build": "Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41", "runtime": "litert-lm", "runtime_version": "0.16.1", "soc": "Broadcom BCM2712", "vendor_sdk": null }, "error": null, "evidence": [ "pi5 LLM sweep row: `litert-lm benchmark --backend cpu --cpu-thread-count 4 -p 256 -d 256 --runs 1 --cache memory` (wave-2 driver pi5_llm_bench.py; --cache memory rather than the house --cache no, which OOM-kills every >=1.2B file on the 8 GB Pi — equivalence measured on granite-350m int8, +2-3%), 3 invocations per file with cool-down to <=52 C between them, vcgencmd measure_temp + get_throttled logged per invocation, peak RSS polled from /proc; throughput = median of the three invocations (spread in metrics); a row counts as measured only when the real-generation gate (`litert-lm run`, degenerate-output check) passed and every invocation exited 0 with get_throttled 0x0", "versions: cpu=Raspberry Pi 5 Model B Rev 1.1, litert-lm=0.16.1, litert-lm-api=0.16.1, platform=Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41, python=3.13.5", "cache mode 'memory'; -p 256 -d 256 --runs 1 --cpu-thread-count 4", "gate ('What is 17 plus 26? Answer with the number only.'): status pass, exit 0, wall 209.8 s, output head \"[thought] Okay, let's see. I need to add 17 and 26 together. Hmm, where do I sta\"", "invocation 0: exit=0 wall_s=398.2 temp 51.0->52.7C throttled=0x0 prefill_tps=18.39 decode_tps=1.96 ttft_s=14.5402 init_s=81.9599 peak_rss_mb=4247", "invocation 1: exit=0 wall_s=400.3 temp 50.5->52.1C throttled=0x0 prefill_tps=19.1 decode_tps=1.97 ttft_s=14.0367 init_s=81.5099 peak_rss_mb=4241", "invocation 2: exit=0 wall_s=392.2 temp 51.0->55.4C throttled=0x0 prefill_tps=20.38 decode_tps=1.97 ttft_s=13.1383 init_s=81.0837 peak_rss_mb=4232" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 256.0, "decode_tps_max": 1.97, "decode_tps_min": 1.96, "init_s": 81.5099, "init_s_max": 81.9599, "init_s_min": 81.0837, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 20.38, "prefill_tps_min": 18.39, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 14.5402, "ttft_s_min": 13.1383 }, "output_match": null, "peak_mem_mb": 4247.0, "prefill_tokens_per_s": 19.1, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 14036.7 }, "source": "data/device_runs/0.16.1/2026-09-01/smollm3-3b-int8__raspberry-pi-5.json" } ] }, "model": { "family": "smollm3", "id": "smollm3-3b-int8", "license": "apache-2.0", "source_url": "https://huggingface.co/litert-community/SmolLM3-3B", "task": "text-generation" }, "pitfalls": [ "Galaxy S26 GPU per the README: runs, 2930/2930 ops across 2 subgraphs on LiteRT GPU (704 MB); Raspberry Pi 5 CPU 19.1 tok/s prefill / 2.0 decode (README tables); this repo's Pi 5 row is the wave-2 journal (device runs).", "The tokenizer in SmolLM3-3B.litertlm was a SentencePiece conversion of the model's BPE tokenizer that lost byte-level semantics for standalone accented letters/symbols, and both bundles later received a metadata-only repair declaring the reasoning channel (thought, …) and a default-system-turn change (every section except LlmMetadata byte-identical, sha256 per section) — read the README's 'Metadata-only change' notes for the exact state of the file you download (README).", "License Apache-2.0, inherited from HuggingFaceTB/SmolLM3-3B (README License).", "LiteRT-LM .litertlm bundle: LiteRT.js cannot run it, so delegation stays null and there is no browser block; device rows come from data/device_runs/." ], "schema_version": "1.2" }